feature: MPL_gpu_imemcpy is not implemented in CUDA

Open dycz0fx opened this issue 2 years ago • 3 comments

In PR https://github.com/pmodels/mpich/pull/6451, ipc read bcast and alltoall requires MPL_gpu_imemcpy to move the data. But this function is only implemented in mpl_gpu_ze.c (no implementation of this function in CUDA or HIP), causing the tests for ipc read bcast and alltoall (alltoall_gpu.c and bcast_gpu.c in test/mpi/coll) to fail.

Aug 31 '23 16:08 dycz0fx

In PR https://github.com/pmodels/mpich/pull/6733, ipc read allgather and allgatherv also requires MPL_gpu_imemcpy

Mar 28 '24 17:03 dycz0fx

Will someone from Intel implement the functionality?

Mar 28 '24 21:03 hzhou

I am not sure. Let me check with Maria.

Mar 28 '24 21:03 dycz0fx